We consider it to be best practice to cleanly separate the parameters of a model from the model itself. At the least, parameters should be defined in a separate section at the start of a file. Ideally, they should be defined in a separate file entirely. This makes version control easier, since the model code typically changes less often than the parameters, and makes it easier to track a simulation project, since the parameter sets can be stored in a database, displayed in a GUI, etc.
At their simplest, individual parameters consist of a name and a value. The value is either a simple type such as a numerical value or a string, or an aggregate of such simple types, such as a set, list or array.
However, we may also wish to specify the physical dimensions of the parameter, i.e., its units, and the range of permissible values.
It is also often useful to specify an object that generates numerical values or strings, such as a random number generator, and treat that object as the parameter.
To support all these uses, we define the Parameter and ParameterRange classes, and various subclasses of the ParameterDist abstract class, such as GammaDist, NormalDist and UniformDist.
Here are some examples of creating Parameter objects:
>>> i1 = Parameter(3)
>>> f1 = Parameter(6.2)
>>> f2 = Parameter(-65.3, "mV")
>>> s1 = Parameter("hello", name="message_to_the_world")
The parameter name, units, value and type can be accessed as attributes:
>>> i1.value
3
>>> f1.type
<type 'float'>
>>> f2.units
'mV'
>>> s1.name
'message_to_the_world'
Parameter objects are not hugely useful at the moment. The units are not used for checking dimensional consistency, for example, and Parameter objects are not drop-in replacements for numerical values - you must always use the value attribute to access the value, whereas it might be nice to define, for example, a class IntegerParameter which was a subclass of the built-in int type.
When investigating the behaviour of a model or in doing sensitivity analysis, it is often useful to run a model several times using a different value for a certain parameter each time (also see the iter_range_keys() and similar methods of the ParameterSet class, below). The ParameterRange class supports this. Some usage examples:
>>> tau_m_range = ParameterRange([10.0, 15.0, 20.0], "ms", "tau_m")
>>> tau_m_range.name
'tau_m'
>>> tau_m_range.next()
10.0
>>> tau_m_range.next()
15.0
>>> [2*tau_m for tau_m in tau_m_range]
[20.0, 30.0, 40.0]
As with taking parameter values from a series or range, it is often useful to pick values from a particular random distribution. Three classes are available: UniformDist, GammaDist and NormalDist. Examples:
>>> ud = UniformDist(min=-1.0, max=1.0)
>>> gd = GammaDist(mean=0.5, std=1.0)
>>> nd = NormalDist(mean=-70, std=5.0)
>>> ud.next()
array([-0.56342352])
>>> gd.next(3)
array([ 0.04061142, 0.05550265, 0.23469344])
>>> nd.next(2)
array([-76.18506715, -68.71229944])
[Note that very similar functionality is available with the RandomDistribution class in the pyNN.random module. We should look at the best ways to avoid duplication].
A problem with parameter sets for large-scale, detailed models is that the list of parameters gets very long and unwieldy, and due to the typically hierarchical nature of such models, the individual parameter names can also get very long, e.g., v1_layer5_pyramidal_apical_dend_gbar_na.
A solution to this is to give the parameter set a hierarchical structure as well, which allows the top-level list of parameters to be very short (e.g. v1, retina and lgn for a visual system simulation) since the top-level parameters are themselves parameter sets.
The simplest way to implement this in Python is using nested dicts. One disadvantage of this is that accessing deeply-nested parameters can be very verbose, e.g. v1['layer5']['pyramidal']['apical_dend']['na']['gbar']. A second disadvantage is that it is tedious to flatten the hierarchy when this becomes necessary, e.g. for serialisation - writing to file, etc.
For these reasons we have created a ParameterSet class, which:
- allows a more convenient notation;
2. enables subsets of the parameters, lower in the hierarchy, to be passed around by themselves;
3. provides convenient methods for reading from/writing to file and for determining the differences between two different parameter sets.
An example of the notation is v1.layer5.pyramidal.apical_dend.na.gbar, which requires only a single . for each level in the hierarchy rather than two “'“s, a “[” and a “]”. This is not much shorter than v1_layer5_pyramidal_apical_dend_gbar_na - the difference is that v1.layer5.pyramidal is itself a ParameterSet object that can be passed as an argument to the pyramidal cell object, which doesn’t care about v1.layer4.spinystellate, let alone retina.ganglioncell.magno.tau_m (while v1_layer5_pyramidal is just a NameError).
ParameterSet objects may be created from a dict:
>>> sim_params = ParameterSet({'dt': 0.11, 'tstop': 1000.0})
or loaded from a URL:
>>> exc_cell_params = ParameterSet("https://neuralensemble.org/svn/NeuroTools/trunk/doc/example.param")
They may be nested:
>>> inh_cell_params = ParameterSet({'tau_m': 15.0, 'cm': 0.5})
>>> network_params = ParameterSet({'excitatory_cells': exc_cell_params, 'inhibitory_cells': inh_cell_params})
>>> P = ParameterSet({'sim': sim_params, 'network': network_params}, label="my_params")
Note that although we show here only numerical parameter values, Parameter, ParameterRange and ParameterDist objects, as well as strings, may also be parameter values.
To see the entire parameter set at once, nicely formatted use the pretty() method:
>>> print P.pretty()
{
"network": {
"excitatory_cells": url("https://neuralensemble.org/svn/NeuroTools/trunk/doc/example.param")
"inhibitory_cells": {
"tau_m": 15.0,
"cm": 0.75,
},
},
"sim": {
"tstop": 1000.0,
"dt": 0.11,
},
}
By default, if the ParameterSet contains other ParameterSets that were loaded from URLs, these will be represented with a url() function in the output, but there is also the option to expand all URLs and show the full contents:
>>> print P.pretty(expand_urls=True)
{
"network": {
"excitatory_cells": {
"tau_refrac": 0.11,
"tau_m": 10.0,
"cm": 0.25,
"synI": {
"tau": 10.0,
"E": -75.0,
},
"synE": {
"tau": 1.5,
"E": 0.0,
},
"v_thresh": -57.0,
"v_reset": -70.0,
"v_rest": -70.0,
},
"inhibitory_cells": {
"tau_m": 15.0,
"cm": 0.75,
},
},
"sim": {
"tstop": 1000.0,
"dt": 0.11,
},
}
If a ParameterSet was loaded from a URL, it may be modified then saved back to the same URL, provided the protocol supports writing:
>>> exc_cell_params.save()
Traceback (most recent call last):
File "<stdin>", line 1, in ?
File "parameters.py", line 266, in save
raise Exception("Saving using the %s protocol is not implemented" % scheme)
Exception: Saving using the https protocol is not implemented
or saved to a different URL:
>>> exc_cell_params.save(url="file:///tmp/exc_params")
The file format is the same as that produced by the pretty() method.
A ParameterSet can be used simply as a dictionary, but can also be converted explicitly to a dict if required:
>>> print sim_params.as_dict()
{'tstop': 1000.0, 'dt': 0.11}
[need to say something about tree_copy()]
There are several different ways to iterate over all or part of the ParameterSet object. keys(), values() and items() work as for dict``s. For the sake of more readable code, ``names() is provided as an alias for keys() and parameters() as an alias for items():
>>> P.names()
['network', 'sim']
>>> exc_cell_params.parameters()
[('tau_refrac', 0.11), ('tau_m', 10.0), ('cm', 0.25),
('synI', {'tau': 10.0, 'E': -75.0}), ('synE', {'tau': 1.5, 'E': 0.0}),
('v_thresh', -57.0), ('v_reset', -70.0), ('v_rest', -70.0)]
To flatten nested parameter sets, i.e., the iterate recursively over all branches of the tree, the the flatten() method returns a dict with keys created by joining the names at each hierarchical level with a separator character (‘.’ by default):
>>> network_params.flatten()
{'excitatory_cells.synI.E': -75.0, 'excitatory_cells.v_rest': -70.0,
'excitatory_cells.tau_refrac': 0.11, 'excitatory_cells.v_reset': -70.0,
'excitatory_cells.v_thresh': -57.0, 'excitatory_cells.tau_m': 10.0,
'excitatory_cells.synI.tau': 10.0, 'excitatory_cells.cm': 0.25,
'inhibitory_cells.cm': 0.75, 'excitatory_cells.synE.tau': 1.5,
'excitatory_cells.synE.E': 0.0, 'inhibitory_cells.tau_m': 15.0}
while the flat() method returns a generator which yields (name, value) tuples.:
>>> for x in network_params.flat():
... print x
...
The ParameterSpace class is a subclass of ParameterSet that is allowed to contain ParameterRange and ParameterDist objects as parameters. This turns the single point in parameter space represented by a ParameterSet into a set of points. For example, the following definition creates a set of six points in parameter space, which can be obtained in turn using the iter_inner() method:
>>> PS = ParameterSpace({
... 'x': 999,
... 'y': ParameterRange([10, 20]),
... 'z': ParameterRange([-1, 0, 1])
... })
>>> for P in PS.iter_inner():
... print P
{'y': 10, 'x': 999, 'z': -1}
{'y': 20, 'x': 999, 'z': -1}
{'y': 10, 'x': 999, 'z': 0}
{'y': 20, 'x': 999, 'z': 0}
{'y': 10, 'x': 999, 'z': 1}
{'y': 20, 'x': 999, 'z': 1}
Putting parameter distribution objects inside a ParameterSpace allows an essentially infinite number of points to be generated:
>>> PS2 = ParameterSpace({
... 'x': UniformDist(min=-1.0, max=1.0),
... 'y': GammaDist(mean=0.5, std=1.0),
... 'z': NormalDist(mean=-70, std=5.0)
... })
>>> for P in PS2.realize_dists(n=3):
... print P
{'y': 1.81311773668, 'x': 0.883293989399, 'z': -73.5871002759}
{'y': 0.299391158731, 'x': 0.371474054049, 'z': -68.6936045978}
{'y': 2.90108202422, 'x': -0.388218831787, 'z': -68.6681724449}
A module for dealing with model parameters.
Parameter ParameterRange - for specifying a list of possible values for a given parameter. ParameterReference - specify a parameter in terms of the value of another parameter. ParameterSet - for representing/managing hierarchical parameter sets. ParameterTable - a sub-class of ParameterSet that can represent a table of parameters. ParameterSpace - a collection of ParameterSets, representing multiple points in
parameter space.
Imported from NeuroTools.parameters.validators
ValidationError - The Exception raised when validation fails
SchemaBase - The base class of all “active” Schema objects to be placed in a ParameterSchema. -> Sublass - Validates the same-path ParameterSet value if it is of the specified type. -> Eval - Validates the same-path ParameterSet value if the provided expression
evaluates (“eval”) to True.
nesteddictwalk - Walk a nested dict structure, using a generator. nesteddictflatten - Return a flattened version of a nested dict structure. string_table - Convert a table written as a multi-line string into a dict of dicts.
validators - A module implementing validation of ParameterSets against ParameterSchema.
A class for specifying a list of possible values for a given parameter.
The value must be an iterable. It acts like a Parameter, but .next() can be called to iterate through the values
This class provides a place-holder for a reference parameter that will later be replaced with the value of the parameter pointed to by the reference. This class also allows for lazy application of operations, meaning that one can use the reference in simple formulas that will get evaluated at the moment the reference is replaced.
Check below which operations are supported.
A class to manage hierarchical parameter sets.
Usage example:
>>> sim_params = ParameterSet({'dt': 0.1, 'tstop': 1000.0})
>>> exc_cell_params = ParameterSet("http://neuralensemble.org/svn/NeuroTools/example.params")
>>> inh_cell_params = ParameterSet({'tau_m': 15.0, 'cm': 0.5})
>>> network_params = ParameterSet({'excitatory_cells': exc_cell_params, 'inhibitory_cells': inh_cell_params})
>>> P = ParameterSet({'sim': sim_params, 'network': network_params})
>>> P.sim.dt
0.1
>>> P.network.inhibitory_cells.tau_m
15.0
>>> print P.pretty()
Like __setitem__, but it will add ParameterSet({}) objects into the namespace tree if needed.
Return a unicode string representing the structure of the ParameterSet. evaluating the string should recreate the object.
ParameterSet definition s should be a Python dict definition string, containing objects of types int, float, str, list, dict plus the classes defined in this module, Parameter, ParameterRange, etc. No other object types are allowed, except the function url(‘some_url’) or ref(‘point.delimited.path’), e.g.:
{ 'a' : {'A': 3, 'B': 4},
'b' : [1,2,3],
'c' : 'hello world',
'd' : url('http://example.com/my_cool_parameter_set')
'e' : ref('level1_param_name.level2_param_name.level3_param_name') }
This is largely the JSON (www.json.org) format, but with extra keywords in the namespace such as ParameterRange, GammaDist, etc.
This expects its arguments to be in the form path=value, where path is a . (dot) delimited path to a parameter in the parameter tree rooted in this ParameterSet instance.
This function replaces the values of each parameter in the args with the corresponding values supplied in the arguments.
Write the parameter set to a text file.
The text file syntax is open to discussion. My idea is that it should be valid Python code, preferably importable as a module.
If url is None, try to save to self._url (if it is not None), otherwise save to url.
A collection of ParameterSets, representing multiple points in parameter space. Created by putting ParameterRange and/or ParameterDist objects within a ParameterSet.
Return a dict with the keys and values of the parameters with ParameterRanges
Example:
>>> p = ParameterSpace({})
>>> p.b = ParameterRange([1,2,3])
>>> p.a = ParameterRange(['p','y','t','h','o','n'])
>>> data = p.get_ranges_values()
>>> data
{'a': ['p', 'y', 't', 'h', 'o', 'n'], 'b': [1, 2, 3]}
An iterator of the ParameterSpace which yields ParameterSets with all combinations of ParameterRange elements
An iterator of the ParameterSpace which yields ParameterSets with all combinations of ParameterRange elements which are given by the keys list.
Note: each newly yielded value is one and the same object so storing the returned values results in a collection of many of the lastly yielded object.
copy=True causes each yielded object to be a newly created object, but be careful because this is spawning many dictionaries!
An iterator of the ParameterSpace which yields the ParameterSet with the ParameterRange given by range_key replaced with each of its values
Return the number of ParameterSets that will be returned by the iter_inner() method.
Return the dimensions and labels of the keys for those elements which are ParameterRanges. range_keys are sorted to ensure the same ordering each time.
Return the index of the current experiment in the dimension of the parameter space i.e. parameter space dimension: [2,3] i.e. index: (1,0)
Example:
p = ParameterSet({})
p.b = ParameterRange([1,2,3])
p.a = ParameterRange(['p','y','t','h','o','n'])
results_dim, results_label = p.parameter_space_dimension_labels()
results = numpy.empty(results_dim)
for experiment in p.iter_inner():
index = p.parameter_space_index(experiment)
results[index] = 2.
A sub-class of ParameterSet that can represent a table of parameters.
i.e., it is limited to one-level of nesting, and each sub-dict must have the same keys. In addition to the possible initialisers for ParameterSet, a ParameterTable can be initialised from a multi-line string, e.g.:
>>> pt = ParameterTable('''
... # col1 col2 col3
... row1 1 2 3
... row2 4 5 6
... row3 7 8 9
... ''')
>>> pt.row2.col3
6.0
>>> pt.column('col1')
{'row1': 1.0, 'row2': 4.0, 'row3': 7.0}
>>> pt.transpose().col3.row2
6.0
This is a function that should be used to load a ParameterSet from a url.
modified_parameters should be a dictionary of parameters and their values. These will be replaced in the loaded parameter set before the references are expanded.
Return a flattened version of a nested dict structure.
Composite keys are created by joining each key to the key of the parent dict using separator.